@vib-rato/ai 0.17.2 → 0.18.0

package/CHANGELOG.md

@@ -2,6 +2,15 @@
 
 ## [Unreleased]
 
+## [0.18.0] - 2026-09-08
+
+- Transport failure facts now retain the Vibrato Usage Gateway's `x-vug-daily-limit`, `x-vug-daily-used`, `x-vug-daily-remaining`, `x-vug-daily-reset`, `x-vug-queue-depth`, `x-vug-inflight`, and `x-vug-queued-ms` headers (raw string values, allowlisted like the existing `retry-after`/`retry-after-ms` entries), and classify a gateway `daily_token_limit` provider code as a `quota` fallback trigger instead of `rate_limit`. Previously the gateway's daily-limit 429 was treated as an ordinary rate limit with a 12-hour retry window, and the `x-vug-daily-*` headers were dropped before reaching consumers. A code-less 429 still classifies as `rate_limit`, and 503 `queue_timeout`/`queue_full` still classify as `server`.
+- OpenAI Codex HTTP failures with an empty response body and empty status text now include the HTTP status in the surfaced message instead of the opaque `Request failed`, while structured provider errors retain their existing messages and classification.
+- OAuth callback responses now serialize provider-controlled result fields as safe JSON script data, preventing callback values from terminating the embedded state element while preserving exact JSON values and callback behavior.
+- Bundled `anthropic/claude-fable-5-1` (Claude Fable 5.1): 1M context, 128K output, `anthropic-adaptive` thinking clamped to `xhigh` on the Messages API, matching the Fable 5 effort ladder exactly. The model is served only by Anthropic's first-party API and requires a Claude Code compatibility version of at least `2.1.251`; older fingerprints are answered with an HTTP 400 `claude_code_version_too_old`, which the `2.1.257` attribution bump clears. Verified live: `2.1.219` returns the 400, `2.1.257` returns 200, `low`/`high`/`xhigh` all stream, and adaptive thinking returns visible summarized content rather than signature-only billed thinking.
+- OpenAI Codex WebSockets are now opt-in on Windows. Bun 1.4.0 can segfault in its Windows TLS WebSocket handshake, so the automatic model preference uses SSE on Windows while the explicit WebSocket setting remains available for operators who have a newer safe runtime.
+## [0.17.3] - 2026-09-03
+
 ## [0.17.2] - 2026-09-03
 
 ## [0.17.1] - 2026-09-03

package/dist/types/providers/anthropic.d.ts

@@ -107,7 +107,7 @@ export interface CpaToolAliasRestoreFailure {
  */
 export declare function parseCpaToolAliasRestoreFailure(error: unknown): CpaToolAliasRestoreFailure | undefined;
 export declare function isCpaToolAliasRestoreFailure(error: unknown): boolean;
-export declare const claudeCodeVersion = "2.1.219";
+export declare const claudeCodeVersion = "2.1.257";
 export declare const claudeCodeEntrypoint = "sdk-cli";
 export declare const claudeToolPrefix: string;
 export declare const claudeCodeSystemInstruction = "You are a Claude agent, built on Anthropic's Claude Agent SDK.";

package/dist/types/providers/openai-codex-responses.d.ts

@@ -25,6 +25,12 @@ export declare function normalizeCodexToolChoice(choice: ToolChoice | undefined,
 export declare function formatCodexUserAgent(platform: string, release: string, arch: string): string;
 export declare const streamOpenAICodexResponses: StreamFunction<"openai-codex-responses">;
 export declare function prewarmOpenAICodexResponses(model: Model<"openai-codex-responses">, options?: Pick<OpenAICodexResponsesOptions, "apiKey" | "headers" | "sessionId" | "signal" | "preferWebsockets" | "providerSessionState">): Promise<void>;
+/**
+ * Bun 1.4.0's Windows WebSocket client can segfault during TLS handshakes.
+ * Keep the model's websocket preference available on other platforms, while
+ * requiring an explicit opt-in on Windows until the bundled runtime is fixed.
+ */
+export declare function isCodexWebSocketSafeByDefault(platform?: NodeJS.Platform): boolean;
 export interface OpenAICodexTransportDetails {
     websocketPreferred: boolean;
     lastTransport?: CodexTransport;

package/dist/types/types.d.ts

@@ -997,6 +997,18 @@ export interface LongContextPricing {
     threshold: number;
     cost: ModelCost;
 }
+/**
+ * Capability facts the serving endpoint advertised for a discovered model
+ * (the `vibrato` object on its OpenAI-style models-list entry). Only fields
+ * that describe what the model can do are carried: a hint can never redirect a
+ * request or change its credentials.
+ */
+export interface DiscoveredModelHint {
+    name?: string;
+    reasoning?: boolean;
+    thinking?: ThinkingConfig;
+    compat?: Pick<OpenAICompat, "supportsReasoningEffort" | "reasoningContentField" | "thinkingFormat">;
+}
 export interface Model<TApi extends Api = any> {
     id: string;
     name: string;
@@ -1022,6 +1034,12 @@ export interface Model<TApi extends Api = any> {
     maxTokens: number;
     /** Explicit models.yml/model-override authority; absent keeps the safe transport default. */
     maxTokensSource?: ModelMaxTokensSource;
+    /**
+     * What the endpoint advertised about this model at discovery time. Kept on
+     * the model so a cached discovery restores it; the registry applies it after
+     * provider overrides and before the user's model overrides.
+     */
+    discoveryHint?: DiscoveredModelHint;
     headers?: Record<string, string>;
     /**
      * Streaming transport override. When `"pi-native"`, `streamSimple` routes

package/dist/types/utils/fallback-transport.d.ts

@@ -35,8 +35,9 @@ export type TransportHeaders = Headers | Record<string, string | undefined>;
  * Structured facts from an upstream HTTP or transport failure. Retry decisions
  * must use these facts rather than provider- or application-owned error text.
  *
- * `headers` is always a plain record limited to the retained retry-signal
- * entries: facts travel on persisted `AssistantMessage`s and through
+ * `headers` is always a plain record limited to the retained retry and
+ * gateway quota/queue signal entries: facts travel on persisted
+ * `AssistantMessage`s and through
  * `structuredClone` snapshots (managed fallback attempt staging), so they must
  * never carry a live `Headers` instance — cloning one throws `DataCloneError`
  * ("The object can not be cloned.") and masks the real provider failure.

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@vib-rato/ai",
-	"version": "0.17.2",
+	"version": "0.18.0",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://github.com/Keonho-Chu/Vibrato",
 	"author": "Yeachan-Heo and Vibrato Contributors",
@@ -40,8 +40,8 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "^0.94.0",
 		"@bufbuild/protobuf": "^2.12.0",
-		"@vib-rato/natives": "0.17.2",
-		"@vib-rato/utils": "0.17.2",
+		"@vib-rato/natives": "0.18.0",
+		"@vib-rato/utils": "0.18.0",
 		"openai": "^6.36.0",
 		"partial-json": "^0.1.7",
 		"zod": "4.4.3"

package/src/model-thinking.ts

@@ -230,7 +230,6 @@ export function modelSupportsReasoningControl<TApi extends Api>(
 	resolvedBaseUrl?: string,
 ): boolean {
 	if (!model.reasoning) return false;
-
 	if (model.api === "openai-completions") {
 		const completionsModel = model as ApiModel<"openai-completions">;
 		const explicitSupport = completionsModel.compat?.supportsReasoningEffort;
@@ -346,6 +345,18 @@ export function clampThinkingLevelForModel<TApi extends Api>(
 	if (!modelSupportsReasoningControl(model) || requested === undefined) {
 		return undefined;
 	}
+	if (model.api === "anthropic-messages" && model.provider === "anthropic") {
+		const baseUrl = model.baseUrl;
+		try {
+			const url = new URL(baseUrl || "https://api.anthropic.com");
+			if (url.protocol !== "https:" || url.hostname !== "api.anthropic.com") return undefined;
+		} catch {
+			return undefined;
+		}
+	}
+	if (!model.thinking) {
+		return undefined;
+	}
 
 	const levels = getSupportedEfforts(model);
 	if (levels.includes(requested)) {

package/src/models.json

@@ -3853,6 +3853,31 @@
 				"maxLevel": "xhigh"
 			}
 		},
+		"claude-fable-5-1": {
+			"id": "claude-fable-5-1",
+			"name": "Anthropic Fable 5.1",
+			"api": "anthropic-messages",
+			"provider": "anthropic",
+			"baseUrl": "https://api.anthropic.com",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 10,
+				"output": 50,
+				"cacheRead": 0.25,
+				"cacheWrite": 12.5
+			},
+			"contextWindow": 1000000,
+			"maxTokens": 128000,
+			"thinking": {
+				"mode": "anthropic-adaptive",
+				"minLevel": "minimal",
+				"maxLevel": "xhigh"
+			}
+		},
 		"claude-haiku-4-5": {
 			"id": "claude-haiku-4-5",
 			"name": "Anthropic Haiku 4.5 (latest)",

package/src/providers/anthropic.d.ts

@@ -107,7 +107,7 @@ export interface CpaToolAliasRestoreFailure {
  */
 export declare function parseCpaToolAliasRestoreFailure(error: unknown): CpaToolAliasRestoreFailure | undefined;
 export declare function isCpaToolAliasRestoreFailure(error: unknown): boolean;
-export declare const claudeCodeVersion = "2.1.219";
+export declare const claudeCodeVersion = "2.1.257";
 export declare const claudeCodeEntrypoint = "sdk-cli";
 export declare const claudeToolPrefix: string;
 export declare const claudeCodeSystemInstruction = "You are a Claude agent, built on Anthropic's Claude Agent SDK.";

package/src/providers/anthropic.ts

@@ -828,7 +828,7 @@ function getCacheControl(
 }
 
 // Stealth mode: Mimic Anthropic Code headers and tool prefixing.
-export const claudeCodeVersion = "2.1.219";
+export const claudeCodeVersion = "2.1.257";
 export const claudeCodeEntrypoint = "sdk-cli";
 export const claudeToolPrefix: string = "proxy_";
 export const claudeCodeSystemInstruction = "You are a Claude agent, built on Anthropic's Claude Agent SDK.";

package/src/providers/openai-codex/response-handler.ts

@@ -25,7 +25,7 @@ const REQUEST_BLOCKED_MESSAGE_RE = /^\s*request blocked\b/i;
 
 export async function parseCodexError(response: Response): Promise<CodexErrorInfo> {
 	const raw = await response.text();
-	let message = raw || response.statusText || "Request failed";
+	let message = raw || response.statusText || `Codex request failed (HTTP ${response.status})`;
 	let friendlyMessage: string | undefined;
 	let rateLimits: CodexRateLimits | undefined;
 	let code: string | undefined;

package/src/providers/openai-codex-responses.d.ts

@@ -25,6 +25,12 @@ export declare function normalizeCodexToolChoice(choice: ToolChoice | undefined,
 export declare function formatCodexUserAgent(platform: string, release: string, arch: string): string;
 export declare const streamOpenAICodexResponses: StreamFunction<"openai-codex-responses">;
 export declare function prewarmOpenAICodexResponses(model: Model<"openai-codex-responses">, options?: Pick<OpenAICodexResponsesOptions, "apiKey" | "headers" | "sessionId" | "signal" | "preferWebsockets" | "providerSessionState">): Promise<void>;
+/**
+ * Bun 1.4.0's Windows WebSocket client can segfault during TLS handshakes.
+ * Keep the model's websocket preference available on other platforms, while
+ * requiring an explicit opt-in on Windows until the bundled runtime is fixed.
+ */
+export declare function isCodexWebSocketSafeByDefault(platform?: NodeJS.Platform): boolean;
 export interface OpenAICodexTransportDetails {
     websocketPreferred: boolean;
     lastTransport?: CodexTransport;

package/src/providers/openai-codex-responses.ts

@@ -2307,6 +2307,15 @@ function recordCodexWebSocketFailure(state: CodexWebSocketSessionState, activate
 	}
 }
 
+/**
+ * Bun 1.4.0's Windows WebSocket client can segfault during TLS handshakes.
+ * Keep the model's websocket preference available on other platforms, while
+ * requiring an explicit opt-in on Windows until the bundled runtime is fixed.
+ */
+export function isCodexWebSocketSafeByDefault(platform: NodeJS.Platform = process.platform): boolean {
+	return platform !== "win32";
+}
+
 function shouldUseCodexWebSocket(
 	model: Model<"openai-codex-responses">,
 	state: CodexWebSocketSessionState | undefined,
@@ -2314,7 +2323,11 @@ function shouldUseCodexWebSocket(
 ): boolean {
 	if (!state || state.disableWebsocket) return false;
 	if (preferWebsockets === false) return false;
-	return isCodexWebSocketEnvEnabled() || preferWebsockets === true || model.preferWebsockets === true;
+	return (
+		isCodexWebSocketEnvEnabled() ||
+		preferWebsockets === true ||
+		(isCodexWebSocketSafeByDefault() && model.preferWebsockets === true)
+	);
 }
 
 export interface OpenAICodexTransportDetails {
@@ -2372,7 +2385,9 @@ export function getOpenAICodexTransportDetails(
 	const websocketPreferred =
 		options?.preferWebsockets === false
 			? false
-			: isCodexWebSocketEnvEnabled() || options?.preferWebsockets === true || model.preferWebsockets === true;
+			: isCodexWebSocketEnvEnabled() ||
+				options?.preferWebsockets === true ||
+				(isCodexWebSocketSafeByDefault() && model.preferWebsockets === true);
 	const state = getCodexWebSocketStateForPublicSession(model, options);
 
 	return {

package/src/providers/pi-native-client.ts

@@ -49,6 +49,29 @@ const NON_WIRE_KEYS = new Set<keyof SimpleStreamOptions>([
 	"fallbackAttempt",
 ]);
 
+/**
+ * Project the caller's {@link Context} onto the wire schema. Runtime tool
+ * objects routinely carry harness-only state (runners, session managers,
+ * fs-stat BigInts) that must never be serialized: BigInt fields make
+ * `JSON.stringify` throw outright, and the rest is dead weight the gateway
+ * re-derives from its own tool registry. Only the protocol-meaningful,
+ * JSON-safe `Tool` fields cross the wire.
+ */
+function buildWireContext(context: Context): Context {
+	if (!context.tools || context.tools.length === 0) return context;
+	return {
+		...context,
+		tools: context.tools.map(tool => ({
+			name: tool.name,
+			description: tool.description,
+			parameters: tool.parameters,
+			...(tool.strict !== undefined ? { strict: tool.strict } : {}),
+			...(tool.customFormat !== undefined ? { customFormat: tool.customFormat } : {}),
+			...(tool.customWireName !== undefined ? { customWireName: tool.customWireName } : {}),
+		})),
+	};
+}
+
 function buildWireOptions(options: SimpleStreamOptions | undefined): Record<string, unknown> {
 	if (!options) return {};
 	const wire: Record<string, unknown> = {};
@@ -166,7 +189,7 @@ export function streamPiNative<TApi extends Api>(
 			const headers = buildHeaders(model as Model<Api>, options?.apiKey);
 			const body = JSON.stringify({
 				modelId: model.id,
-				context,
+				context: buildWireContext(context),
 				options: buildWireOptions(options),
 				stream: true,
 			});

package/src/types.d.ts

@@ -997,6 +997,18 @@ export interface LongContextPricing {
     threshold: number;
     cost: ModelCost;
 }
+/**
+ * Capability facts the serving endpoint advertised for a discovered model
+ * (the `vibrato` object on its OpenAI-style models-list entry). Only fields
+ * that describe what the model can do are carried: a hint can never redirect a
+ * request or change its credentials.
+ */
+export interface DiscoveredModelHint {
+    name?: string;
+    reasoning?: boolean;
+    thinking?: ThinkingConfig;
+    compat?: Pick<OpenAICompat, "supportsReasoningEffort" | "reasoningContentField" | "thinkingFormat">;
+}
 export interface Model<TApi extends Api = any> {
     id: string;
     name: string;
@@ -1022,6 +1034,12 @@ export interface Model<TApi extends Api = any> {
     maxTokens: number;
     /** Explicit models.yml/model-override authority; absent keeps the safe transport default. */
     maxTokensSource?: ModelMaxTokensSource;
+    /**
+     * What the endpoint advertised about this model at discovery time. Kept on
+     * the model so a cached discovery restores it; the registry applies it after
+     * provider overrides and before the user's model overrides.
+     */
+    discoveryHint?: DiscoveredModelHint;
     headers?: Record<string, string>;
     /**
      * Streaming transport override. When `"pi-native"`, `streamSimple` routes

package/src/types.ts

@@ -1213,6 +1213,19 @@ export interface LongContextPricing {
 	cost: ModelCost;
 }
 
+/**
+ * Capability facts the serving endpoint advertised for a discovered model
+ * (the `vibrato` object on its OpenAI-style models-list entry). Only fields
+ * that describe what the model can do are carried: a hint can never redirect a
+ * request or change its credentials.
+ */
+export interface DiscoveredModelHint {
+	name?: string;
+	reasoning?: boolean;
+	thinking?: ThinkingConfig;
+	compat?: Pick<OpenAICompat, "supportsReasoningEffort" | "reasoningContentField" | "thinkingFormat">;
+}
+
 export interface Model<TApi extends Api = any> {
 	id: string;
 	name: string;
@@ -1238,6 +1251,12 @@ export interface Model<TApi extends Api = any> {
 	maxTokens: number;
 	/** Explicit models.yml/model-override authority; absent keeps the safe transport default. */
 	maxTokensSource?: ModelMaxTokensSource;
+	/**
+	 * What the endpoint advertised about this model at discovery time. Kept on
+	 * the model so a cached discovery restores it; the registry applies it after
+	 * provider overrides and before the user's model overrides.
+	 */
+	discoveryHint?: DiscoveredModelHint;
 	headers?: Record<string, string>;
 	/**
 	 * Streaming transport override. When `"pi-native"`, `streamSimple` routes

package/src/utils/fallback-transport.d.ts

@@ -35,8 +35,9 @@ export type TransportHeaders = Headers | Record<string, string | undefined>;
  * Structured facts from an upstream HTTP or transport failure. Retry decisions
  * must use these facts rather than provider- or application-owned error text.
  *
- * `headers` is always a plain record limited to the retained retry-signal
- * entries: facts travel on persisted `AssistantMessage`s and through
+ * `headers` is always a plain record limited to the retained retry and
+ * gateway quota/queue signal entries: facts travel on persisted
+ * `AssistantMessage`s and through
  * `structuredClone` snapshots (managed fallback attempt staging), so they must
  * never carry a live `Headers` instance — cloning one throws `DataCloneError`
  * ("The object can not be cloned.") and masks the real provider failure.

package/src/utils/fallback-transport.ts

@@ -40,8 +40,9 @@ export type TransportHeaders = Headers | Record<string, string | undefined>;
  * Structured facts from an upstream HTTP or transport failure. Retry decisions
  * must use these facts rather than provider- or application-owned error text.
  *
- * `headers` is always a plain record limited to the retained retry-signal
- * entries: facts travel on persisted `AssistantMessage`s and through
+ * `headers` is always a plain record limited to the retained retry and
+ * gateway quota/queue signal entries: facts travel on persisted
+ * `AssistantMessage`s and through
  * `structuredClone` snapshots (managed fallback attempt staging), so they must
  * never carry a live `Headers` instance — cloning one throws `DataCloneError`
  * ("The object can not be cloned.") and masks the real provider failure.
@@ -147,14 +148,29 @@ function stringValue(value: unknown): string | undefined {
 	return typeof value === "string" ? value : undefined;
 }
 
-/** Retry-signal headers retained on transport facts; everything else is dropped. */
-const RETAINED_TRANSPORT_HEADERS = ["retry-after", "retry-after-ms"] as const;
+/**
+ * Retry-signal headers retained on transport facts; everything else is
+ * dropped. The `x-vug-*` entries are the Vibrato Usage Gateway's quota and
+ * queue-depth signals; their values are preserved as raw strings and
+ * interpreted by consumers, not by this module.
+ */
+const RETAINED_TRANSPORT_HEADERS = [
+	"retry-after",
+	"retry-after-ms",
+	"x-vug-daily-limit",
+	"x-vug-daily-used",
+	"x-vug-daily-remaining",
+	"x-vug-daily-reset",
+	"x-vug-queue-depth",
+	"x-vug-inflight",
+	"x-vug-queued-ms",
+] as const;
 
 const RETAINED_TRANSPORT_HEADER_SET: ReadonlySet<string> = new Set(RETAINED_TRANSPORT_HEADERS);
 
 /**
- * Reduce transport headers to the retained retry-signal entries in a plain
- * record, so facts stay structured-cloneable and JSON-serializable and never
+ * Reduce transport headers to the retained retry and gateway quota/queue
+ * signal entries in a plain record, so facts stay structured-cloneable and JSON-serializable and never
  * persist arbitrary response headers into session files.
  *
  * Exception-safe by contract: inspection uses only `Headers.get()` results
@@ -276,9 +292,22 @@ export function transportFailureFacts(
 	};
 }
 
+/**
+ * Exception-safe by contract: `new Headers(record)` throws on an invalid
+ * header value (e.g. a CRLF in a gateway-supplied timestamp like
+ * `x-vug-daily-reset`), and the retained record is otherwise free-form
+ * provider input. A throw here falls through to `undefined` — the same
+ * "no retry hint" result as if no headers were retained at all — instead of
+ * propagating a TypeError out of classification.
+ */
 function headersOf(headers: TransportHeaders | undefined): Headers | undefined {
 	if (headers instanceof Headers) return headers;
-	return headers ? new Headers(headers as Record<string, string>) : undefined;
+	if (!headers) return undefined;
+	try {
+		return new Headers(headers as Record<string, string>);
+	} catch {
+		return undefined;
+	}
 }
 
 function parseRetryAfterSeconds(value: string | null, now = Date.now()): number | undefined {
@@ -305,7 +334,12 @@ function isQuotaCode(code: string | undefined): boolean {
 		code === "quota_exhausted" ||
 		code === "usage_limit_reached" ||
 		code === "usage_not_included" ||
-		code === "out_of_credits"
+		code === "out_of_credits" ||
+		// Vibrato Usage Gateway daily-limit code: a 429 with this code is a
+		// quota exhaustion, not an ordinary rate limit, even though the
+		// gateway's retry-after window can be many hours long. A code-less
+		// 429 from an older gateway still falls through to `rate_limit`.
+		code === "daily_token_limit"
 	);
 }
 

package/src/utils/oauth/callback-server.ts

@@ -19,6 +19,13 @@ const DEFAULT_TIMEOUT = 300_000;
 const DEFAULT_HOSTNAME = "localhost";
 const CALLBACK_PATH = "/callback";
 
+function serializeScriptData(value: unknown): string {
+	return JSON.stringify(value)
+		.replaceAll("<", "\\u003c")
+		.replaceAll("\u2028", "\\u2028")
+		.replaceAll("\u2029", "\\u2029");
+}
+
 export type CallbackResult = { code: string; state: string };
 
 export interface OAuthCallbackFlowOptions {
@@ -259,7 +266,7 @@ export abstract class OAuthCallbackFlow {
 		});
 
 		return new Response(
-			(templateHtml as unknown as string).replaceAll("__OAUTH_STATE__", JSON.stringify(resultState)),
+			(templateHtml as unknown as string).replaceAll("__OAUTH_STATE__", () => serializeScriptData(resultState)),
 			{
 				status: resultState.ok ? 200 : 500,
 				headers: { "Content-Type": "text/html" },